New AI Framework Enhances Underwater Target Recognition

VERA Brief

AI-generated. Grounded in the article and its cited sources.

Researchers created a new AI framework to improve underwater target recognition in synthetic aperture sonar imagery. The framework uses parameter-efficient methods to overcome data scarcity and background clutter challenges.

Key facts

• A new framework adapts DINOv3 Vision Transformer models for automatic target recognition using underwater synthetic aperture sonar data.
• The approach uses Low-Rank Adaptation to bridge natural image pretraining with underwater acoustic propagation.
• Hard-negative mining is employed to improve discrimination against acoustic mimics.
• Supervised Contrastive Learning is used to enhance separation between target and clutter representations.
• Low-Rank Adaptation alone increased the area under the precision-recall curve from 0.300 to 0.679 +/- 0.027 while training only 0.26 percent of weights.